Worked Design of a Group Sequential Trial with Delayed Benefit

From a Delayed-Effect Assumption to a Trial Design

A trial designed around a constant hazard ratio can be poorly matched to a treatment whose benefit is expected only after an induction period, biological latency, or completion of a treatment phase. The design must connect that scientific expectation to the test statistic, the interim schedule, and an interpretable summary of survival. This tutorial develops one hypothetical design as a single chain of decisions. It is not a template for choosing a delay after inspecting trial results: the model, weight, estimand, and monitoring rules must be prespecified.

Suppose adults are randomized equally to control or treatment, and follow-up time \(t\) is measured in months from randomization. A useful planning model assumes an exponential control hazard and a treatment hazard that initially matches control, then falls at a specified onset time. This is a simplifying design model, not a claim that real survival data must be exponential. The weighted logrank analysis can test a broader null of equal survival distributions, while the model supplies a plausible alternative for evaluating power and summaries.

This model assumes a common onset time and a constant hazard within each specified segment. It also assumes that the anticipated pattern is scientifically credible across the relevant population. For the usual survival-score calculations, censoring must be independent of the event process within treatment groups, conditional on any variables used in the analysis; follow-up and event times must be recorded consistently. Randomization supports the treatment comparison but does not make informative censoring harmless.

Worked Design: Model, Test, and Monitoring Plan

Consider a hypothetical two-arm trial with equal allocation. For planning, let the control hazard be \(0.080\) per month, the treatment onset be \(\delta=6\) months, and the post-onset treatment hazard be \(0.052\) per month. Thus \(\theta=0.052/0.080=0.650\). The model says there is no treatment hazard difference through month 6 and a 35% lower instantaneous hazard thereafter. It does not say that every treated participant experiences a 35% longer survival time.

For the primary efficacy analysis, prespecify a one-sided weighted logrank test, with positive values defined to favor treatment. At each event time \(t_j\), let \(d_{1j}\) be the number of treatment-arm events, \(Y_{1j}\) the treatment risk set, and \(d_j,Y_j\) the corresponding totals across both arms. Under the null, the expected treatment events at that time are \(Y_{1j}d_j/Y_j\). Use weight \(w(t)=0\) through month 6 and \(w(t)=1\) after month 6. The weighted score and its null variance are

Here \(Y_{0j}\) is the control risk set. The sign convention makes \(Z_w>0\) evidence of fewer treatment events than expected, conditional on the risk sets. Tied-event implementations may use the corresponding standard tie adjustment. The weight is fixed by time since randomization, so it does not depend on the observed treatment difference. Under the null and regularity conditions, the score is approximately normal; its sequential values are correlated because the looks use accumulating data.

Use three event-driven analyses at 120, 180, and 240 total primary events. Prespecify a one-sided overall type I error of \(0.025\). The actual efficacy boundaries must be calibrated under the null for this weight, event schedule, and chosen analysis procedure; they should not be borrowed from a different monitoring plan. For illustration, suppose null calibration gives standardized boundaries 3.35, 2.40, and 2.02 at the three looks. The design stops for efficacy at the first look with \(Z_w\) at or above its boundary. Any futility rule, including whether stopping for futility is binding, must also be stated in the protocol and included in performance calculations.

RMST as a Complementary Treatment Summary

A weighted logrank test targets evidence against equality of survival distributions using a chosen time weighting. A complementary summary is the restricted mean survival time (RMST) through a prespecified horizon \(\tau\), defined as the area under the survival curve up to that time. It has units of time and can be interpreted without reducing a non-proportional hazard pattern to one constant hazard ratio. The horizon must be chosen in advance and supported by adequate follow-up.

For a constant hazard \(\lambda\), \(S(t)=e^{-\lambda t}\), so the area over an interval can be integrated directly. With a piecewise hazard, survival is continuous at the onset: the post-onset survival starts from the probability of surviving to \(\delta\), not from 1.

Design Justification and Operating Characteristics

The design is defensible only if the choices work together. A delayed-effect alternative makes an ordinary unweighted logrank test potentially less efficient than a test that emphasizes post-onset events. The chosen indicator weight expresses a strong version of that emphasis: events before month 6 contribute no primary weighted score. This may improve sensitivity to the specified delay, but it can lose information if benefit begins earlier, later, or gradually. The test should be selected for a prespecified scientific rationale, not because it gives the most favorable result in the observed data.

A group sequential design must account for repeated looks. Under the null, standardized score statistics at information times \(I_1,\ldots,I_K\) are often approximated by a joint normal process with correlations \(\operatorname{Corr}(Z_k,Z_l)\approx\sqrt{I_k/I_l}\) for \(I_k\leq I_l\). That approximation motivates sequential boundary calibration, but with a weighted statistic and event-driven looks its adequacy should be checked under the planned design. The boundary sequence controls the probability of at least one false efficacy crossing under the null only when calibrated for the actual analysis plan.

In practice, simulate trial histories under both the null and the delayed-effect alternatives. Each simulated trial should reproduce randomization, event and censoring processes, the prespecified weight, look triggers, stopping rules, and analysis method. Estimate type I error as the proportion of null trials crossing an efficacy boundary, and power as the proportion of alternative trials crossing one. The event target can then be adjusted until power meets its target while the calibrated type I error remains controlled. Calendar dates and the number to enroll require separate accrual and follow-up assumptions; event targets alone do not determine those quantities.

Sensitivity, Diagnostics, and Limitations

A design based on one assumed onset and one post-onset hazard ratio should be challenged with alternatives. For example, later onset reduces the period during which the weight captures the treatment difference. A weaker post-onset effect also reduces the expected weighted score. RMST provides a useful additional check: the same trial may have a meaningful 24-month RMST contrast even when a particular weighted test has limited power, or the reverse. These summaries answer related but different questions and should not be presented as interchangeable.

Before interpreting interim data, review whether the planned analysis remains credible: check event counts and timing, risk-set sizes, censoring by arm, data completeness, and whether the observed pattern is broadly compatible with the assumed delayed-effect rationale. A survival-curve pattern that differs from the planning model is a warning to scrutinize interpretation and sensitivity analyses, not a license to change the weight or onset based on the interim treatment difference. Any adaptations must be prespecified and preserve valid error control.

The weighted logrank test and RMST contrast also have limitations. The indicator weight discards pre-onset events from its score and can be inefficient or misleading when the delay is misspecified. RMST depends on the chosen horizon and may be imprecise if few participants remain under observation near that horizon. Neither measure removes the need to report absolute survival estimates, follow-up context, and uncertainty. The design should report the test result and the RMST estimate with an appropriate interval, while avoiding a causal interpretation beyond the randomized treatment contrast and the trial's target population.

• Common mistake: treating the interim weighted \(Z\)-statistic as if it were an ordinary final test at every look. Correction: compare it with its own prespecified sequential boundary and use sequentially adjusted inference.
• Common mistake: interpreting a hazard ratio below 1 as a constant effect throughout follow-up. Correction: in this model the ratio is 1 until onset and only then equals the specified post-onset ratio.
• Common mistake: assuming a chosen delayed-effect weight guarantees power if the effect is delayed. Correction: evaluate power and type I error under plausible delay, effect, censoring, and follow-up scenarios.